Struct Flux2ForwardInput 

pub struct Flux2ForwardInput<'a> {
    pub hidden_states: &'a [f32],
    pub encoder_hidden_states: &'a [f32],
    pub timestep: &'a [f32],
    pub timestep_target: Option<&'a [f32]>,
    pub guidance: Option<&'a [f32]>,
    pub img_ids: &'a [f32],
    pub txt_ids: &'a [f32],
    pub batch: usize,
    pub img_seq: usize,
    pub txt_seq: usize,
}

Inputs for one transformer forward (noise prediction).

Fields

hidden_states: &'a [f32]

Image latents [batch, img_seq, in_channels].

encoder_hidden_states: &'a [f32]

Text encoder states [batch, txt_seq, joint_attention_dim].

timestep: &'a [f32]

Per-batch timestep (sigma); multiplied by 1000 inside the model.

timestep_target: Option<&'a [f32]>

Optional second time for flow-map embedding (average with timestep).

guidance: Option<&'a [f32]>

Per-batch guidance scale; multiplied by 1000 when guidance embeds are enabled.

img_ids: &'a [f32]

Position ids [img_seq + txt_seq, 4] (concatenated txt then img along axis 0).

txt_ids: &'a [f32]

batch: usize

img_seq: usize

txt_seq: usize